AlphaFold: Improved protein structure prediction using potentials from deep learning

15:30-16:15, January 28 @ 1A

Talk/ Overview

Proteins are essential for nearly every process in living organisms and their function is determined by the 3D structure of the component amino acids, making structure determination a grand challenge in biology. Experimental determination is difficult and does not scale to the billions of proteins being discovered through genetic sequencing, so accurate computational structure prediction is essential to advancing biological knowledge. DeepMind's AlphaFold protein structure prediction system was ranked first in free-modelling at the CASP13 (Critical Assessment of Protein Structure Prediction) Biennial blind assessment of protein structure prediction methods. The system relies upon prediction of inter-residue distances by a very deep neural network. Using these distance distributions and a reference distribution from a similar neural network, we construct a potential and show that we can optimize this potential by a simple application of gradient descent, as well as with a more conventional fragment assembly / simulated annealing algorithm. Despite not using templates the system also performed well in the CASP template-based category. We will discuss the training and use of the neural network and present contact- and structure-prediction results from the CASP assessment and indicate potential future directions. 

Talk/ Speakers

Andrew Senior

Research Scientist, DeepMind

AMLD / Global partners